UK: 
USA:
I recently had the opportunity to speak with a senior representative from the FIA from Taiwan.
The Fiscal Information Agency (FIA) is a department under Taiwan’s Ministry of Finance responsible for developing, maintaining, and overseeing the country’s financial data systems and tax-related technology infrastructure. It serves as the central hub for digital transformation efforts, including the design and governance of applications used for tax collection, taxpayer services, and financial auditing.
By consolidating responsibilities under one agency, the FIA ensures that data remains standardized and securely managed across all departments, enabling effective use of technologies like artificial intelligence to improve compliance and operational efficiency.
The conversation provided a fascinating look into how one of Asia’s most tech-forward tax administrations is now looking towards incorporating artificial intelligence, data modelling, and centralised systems to improve tax compliance and detect fraud – having already implemented e-invoicingElectronic invoicing - widely referred to as e-invoicing - is the exchange of a digital document between a supplier and a buyer. E-invoices are issued, transmitted and received in a structured data format that enabled automatic and electronic processing. They contain data in a machine-readable format so that an AP system can read an invoice without manual data entry, leading to faster and more efficient invoicing. years ago, well ahead of many other countries only beginning to explore it now.
Standardising data: The foundation of AI in tax
One of the most critical lessons shared was that successful AI implementation in tax begins with data standardisation. Taiwan’s tax authority emphasised that before AI could deliver any meaningful results, they had to ensure the underlying data was consistent and well-structured.
Without standardised data, it becomes significantly harder to locate, analyse, or leverage information across different systems. The more standardised the data is, the easier it is not only to manage but also to make it AI-ready. This focus on consistency laid the foundation for many of the more advanced AI capabilities they later developed. Having centralised financial data systems provides the ideal platform to really optimise the use of AI.
A centralised data and development model
By adopting a centralised model for both data storage and software development, rather than allowing different departments to build their own disparate tools or manage their own data in silos, they created a unified data centre and development team. This centralisation has led to far better integration and connectivity across systems. The FIA has effectively centralised its data storage and software development through several strategic initiatives:
Unified Information Systems: The FIA oversees the planning, coordination, and evaluation of information systems across the Ministry of Finance (MOF) and its subordinate agencies. This centralised oversight ensures consistency and integration across various platforms.
Integrated Tax Services: Embracing a people-oriented approach, the FIA has led inter-ministerial collaborations to integrate services. Since 2017, they’ve implemented a one-stop service for real estate transfers, consolidating operations of 22 local taxation and land administration departments.
Cloud Infrastructure Development: The FIA has initiated projects like the Cloud Infrastructure Project in Tax Cloud Generation (2021-2025) to modernise and centralize tax-related data storage and processing, enhancing efficiency and scalability. This centralised and standardised approach enables seamless data sharing across departments, allowing AI tools to cross-reference information—such as real estate records—for anomalies. This level of integration empowers the tax authority to detect unusual activity that might otherwise go unnoticed in isolated systems.
Digitalisation of Tax Processes: The FIA has been instrumental in digitising various tax systems, including individual income tax, profit-seeking enterprise income tax, and business tax. This digital transformation centralises data management and streamlines tax administration, allowing for better efficiency, accessibility, and compliance across the board.
To support this transformation, the FIA currently employs over 380 internal engineers and brings in approximately 300 contractors. All applications are built by this centralised team, ensuring uniformity and compatibility across systems. This approach significantly reduces duplication and inefficiencies, and makes it easier to integrate cross-functional features like AI. With all data flowing into a single ecosystem, the FIA is in a much stronger position to analyse it effectively and respond quickly to compliance risks.
AI for tax avoidance detection using graph theory
One of the most compelling aspects of our discussion centered around their use of artificial intelligence to detect tax avoidance. The team employs graph theory, a powerful analytical technique that models relationships between entities – such as individuals, companies, or transactions – as nodes, with their interactions or connections represented as edges.
This method helps uncover hidden patterns and associations that would be difficult to detect using traditional linear methods. Graph theory is widely known and well-documented and so was relatively easy to introduce within Taiwan’s existing systems, although the integration with data and how to get the most out of it was much more of a challenge and the authority had to learn and adapt as they developed the process.
AI models can use graph structures to identify suspicious relationships or patterns that may indicate tax avoidance, including VAT fraud. For example, if multiple companies appear to share directors, bank accounts, or unusual transactional flows, graph theory can map these connections and highlight clusters or paths of behaviour that suggest coordinated activity.
If a business is artificially reducing its profits through circular trading – where companies buy and sell the same goods between themselves at manipulated prices – this would be visualised as a loop or cycle in the graph. Similarly, in VAT carousel fraud, goods may “circle” across borders to exploit refunds, appearing as tight-knit clusters of short-lived firms with rapid, implausible trades.
AI can also integrate diverse data sources to enhance detection: cross-referencing VAT returns with invoices, bank statements, and customs declarations to spot discrepancies; analysing supply chains for abnormal pricing or volumes and tracking temporal patterns like sudden refund spikes.
Shops such as hair salons or nail bars – often flagged in money laundering typologies – could conceptually have their reported revenue compared to average foot traffic in a specific area. If a business shows significantly higher earnings than others with similar customer volumes, this discrepancy could be flagged by AI models as a potential indicator of suspicious financial activity.
The introduction of e-invoicing will provide another level of quality data that can be used in this process. Third-party data – such as company registries, offshore account links, or even a business’s lack of online presence – can enrich the graph, while network analysis of intermediaries like accountants or even known individuals tied to multiple high-risk entities adds another layer of scrutiny.
Cross-border data from systems like the EU’s Transaction Network Analysis (TNA) can further reveal asymmetric tax reporting. These patterns – whether cycles, outliers, or anomalous clusters – are flagged by the AI as potential red flags and escalated for further review by investigators, combining machine precision with human expertise to tackle both simple underreporting and sophisticated fraud schemes.
Whilst a challenge, you can see why the team at the FIA have started this investment in the use of AI and why they are likely to get great results because of the foresight of the Taiwan ministry of finance to create a centralised data centre and team with the FIA, they have the best starting point to be able to move forward.
Multi-level AI analysis and escalation paths
Another revelation is that their AI implementation isn’t an ‘all or nothing’ approach. Instead, it works in levels. At the first level, AI is used to scan for obvious red flags or anomalies, such as unusually low profit declarations or suspicious revenue patterns. If any irregularities are found, they can escalate the investigation to a second or even third level, bringing in additional datasets and applying more complex models.
For example, if a company’s declared revenue doesn’t align with the amounts shown in their bank records, the AI can prompt a deeper analysis. This might involve checking invoice data against actual bank transfers to validate whether the payments match declared income. This layered approach allows them to run quick checks across large datasets without overloading the system or performing deep analysis on every single transaction unnecessarily. It’s only when issues are detected that they dive deeper, sometimes even automating the next step.
Connecting invoices, bank records, and GL data
To cross-validate financial data, a technique similar to the principles of SAF-TSAF-T (Standard Audit File for Tax) is a file type based on the XML standard. It is created in a standard readable format from data exports taken from accounting records. SAF-T is used internationally to ensure the fast and secure digital transfer of tax information. It is known for its high level of security, ability to simplify the collection of tax data and simple readability due to its standardised format. (Standard Audit File for Tax) can be applied. Where SAF-T reporting takes the VAT data reported and reconciles this with the transactional data and then reconciles the transactional data with the GL data, this approach involves consolidating a full spectrum of VAT, transactional data – including invoices, bank transactions, and general ledger entries – into a single framework.
By enabling comparisons across these interconnected sources, tax authorities can more effectively identify inconsistencies such as undeclared income or transactions that bypass standard tax reporting channels.
This concept ensures that data is not only accurate but also properly entered into the system. For instance, it helps flag cases where a transaction might be posted directly to the general ledger without passing through the appropriate tax engine, potentially obscuring its visibility for tax reporting.
When discrepancies emerge, this cross-referencing method can trace anomalies through the complete financial flow, offering clear insights into where and how issues originate. This highlights the critical importance of maintaining high-quality master data and establishing a verifiable digital link between reported tax data and its original source. Without this connection, it becomes difficult to trace, validate, or audit the integrity of the information being reported.
The broader context: Why other tax authorities struggle with AI
Although Taiwan and some other countries have made considerable progress, it’s clear that many other tax authorities around the world are far behind in their adoption of AI. One of the key barriers is capability – many tax departments simply don’t have the internal skills or technological infrastructure to support AI-driven workflows.
However, perhaps more significantly, it’s the quality of the data and the way it’s collected that limits AI’s effectiveness in most jurisdictions. Without clean, structured, and centralised data, any AI model is bound to produce unreliable results, or worse, miss fraudulent activity entirely.
A second key barrier lies in the influence of legal and privacy frameworks – often shaped by tax lawyers – which can limit the full potential of data use in tax administration. While protecting personal data is crucial and the laws enforced, these restrictions sometimes hinder progress in leveraging digital tools and optimising processes.
Rather than defaulting to a ‘balance’ between privacy and utility, a ‘tax-first’ mindset should guide innovation. Practical solutions – such as masking personal data while still allowing access to non-personal corporate data – can provide safeguards without obstructing data-driven reform.
Cultural differences between countries may significantly influence the advancement of tax authorities. In Taiwan, the tax authority is held in very high regard, with its initiatives taken seriously by both businesses and the public. When the tax authority identifies a need for change, it is typically prioritised.
This respect and responsiveness enable swift implementation of reforms and foster innovation. The Fiscal Information Agency (FIA) exemplifies this approach, demonstrating how cultural trust in public institutions can directly support digital transformation.
In contrast, in countries like Germany, the emphasis is placed more heavily on individual privacy (combination of the EU’s General Data Protection Regulation (GDPR), and its own Federal Data Protection Act (BDSG) and the Telecommunications-Telemedia Data Protection Act (TTDSG)), which can lead to stricter limitations on data usage and potentially slower progress in digital tax initiatives.
Cultural differences between countries significantly influence the advancement of tax authorities heavily dictating how much and how far they can go, proving that indeed, ‘culture eats strategy for breakfast’!
In Taiwan, the tax authority enjoys immense public respect, with its initiatives taken seriously by businesses and citizens alike. This cultural reverence creates an environment where calls for reform are prioritised swiftly, enabling rapid implementation and fostering innovation that is far more powerful than the best strategy that can never be implemented!
XML and the push for standardised data exchange
In line with its emphasis on data quality and standardisation, Taiwan’s tax authority has also moved toward a standardised data exchange format, specifically using XMLExtensible Markup Language is a markup language and file format for storing, transmitting, and reconstructing arbitrary data..
This choice isn’t arbitrary – it enables interoperability between different systems and ensures that data is being captured and interpreted consistently. By enforcing XML as the exchange format, it can validate input from taxpayers and ensure that information adheres to a uniform schema before it’s ever ingested by systems.
It’s a practical and forward-looking decision that lays the groundwork for cross-platform automation.
Could AI learn from past fraud cases?
During the meeting, I asked whether the FIA had explored the use of AI not just to detect anomalies, but to learn from known cases of tax fraud. For instance, if a company is caught committing fraud – perhaps through whistleblowing or investigative work outside of AI – could that company’s data be used to train a model? The idea is to feed the AI not with abstract rules, but with concrete examples of what fraudulent data actually looks like, enabling it to spot similar cases in the future.
The response was nuanced. Technically, it acknowledged this kind of machine learning is possible, and in theory, it could be highly effective.
However, they also noted that such an approach would be legally complex, especially concerning privacy and evidentiary standards. Implementing AI models trained on real-world fraud would require strong governance and possibly new legislation and that it would be very complex to set up due to the huge volumes of data and complexities. For now, it hasn’t moved forward with this type of training, but it hasn’t ruled it out either.
AI-powered support and the risk of hallucination
The conversation also turned to how AI might be used to enhance taxpayer support, such as helplines or digital guidance portals. This is something that the Spanish tax authorities did with great results reducing the number of calls to the tax authority by 90% because people were able to use the tax authority chatbot to find the answers to their questions instead.
However, they were quick to point out a major challenge: AI hallucination. AI hallucination refers to when an artificial intelligence model generates outputs that are incorrect, fabricated, or unsupported by its training data or the input provided, yet presents them with confidence as if they were true.
It’s like the AI “imagining” something that doesn’t align with reality, often because it’s trying to fill gaps, extrapolate beyond its knowledge, or mimic patterns it’s learned without verifying facts. Even the most sophisticated language models can occasionally generate plausible-sounding but entirely incorrect information. In a legal or tax context, relying on such advice could have serious repercussions.
To mitigate this risk, I recommended that any use of AI in decision-making be accompanied by a recorded log of both the queries submitted and the answers received.
This decision log could serve as an auditable trail in case a decision is ever challenged. Such a log could include the date, the question you asked, the name of the product used such as a tax authority chatbot, and the final advice selected. It would provide the kind of transparency and traceability that regulators increasingly expect.
Unified invoice system and receipt lotteries
One of the most innovative mechanisms used in Taiwan is the Unified Invoicing System, a digital infrastructure that assigns unique identification numbers to both invoices and receipts. This system treats receipts – such as those issued in shops or restaurants – not as informal proof of payment but as official tax documents. Each receipt includes a government-issued number that can be tracked.
To encourage consumers to demand proper receipts (and thus improve tax compliance), Taiwan has cleverly linked these invoice numbers to a national lottery. Every receipt serves as a lottery ticket. Instead of a traditional draw system where only issued tickets can win, the Taiwanese model allows users to match their receipt numbers against winning numbers. Even partial matches can yield small prizes, incentivising widespread participation and maximizing the number of receipts collected and retained. It’s a textbook example of using behavioural economics – a carrot rather than a stick – to drive better compliance.
The scale of this initiative is staggering. Around 11 billion unique invoice numbers are allocated to businesses each year.
Rather than generating these numbers in real time, companies are issued large blocks of numbers based on their expected transaction volume. They then incorporate these into their own internal invoice numbers, ensuring consistency and traceability without compromising operational flexibility.
This requirement can pose challenges for global firms accustomed to using their own internal invoice numbering schemes. Being mandated to adopt an unbroken sequential numbering system can initially seem like a significant operational change. In reality, it is often a straightforward adjustment, but for companies unfamiliar with this approach, it may appear complex simply because it diverges from their established global practices.
The Covid pandemic unexpectedly accelerated the adoption of digital invoicing. Concerns about surface transmission and the handling of paper documents led many businesses to explore contactless, electronic alternatives. Taiwan capitalised on this shift, encouraging companies to adopt electronic invoicing as a safer and more efficient alternative. As a result, 87% of invoices are now electronic, with efforts continuing to bring the remaining 13% into the fold. COVID-19, while disruptive, became a springboard for deeper digital transformation. A culture of assistance over punishment
Use the tools for the greater good
One of the defining characteristics of Taiwan’s tax administration is its supportive approach to compliance. The assumption is not that taxpayers are trying to cheat, but that they are generally trying to get things right.
So, when discrepancies are found – whether through AI analysis, audit processes, or data cross-checks – the first step is not to issue a penalty, but to notify the taxpayer and allow them to correct the issue. Only in cases of repeated offenses or clear negligence do fines come into play.
This ethos reinforces a collaborative relationship between taxpayer and tax authority. It encourages openness, responsiveness, and better data sharing, all of which feed directly into the success of digital initiatives like e-invoicing and AI-driven audit tools. It’s a clear contrast to many systems around the world where the stick rather than the carrot is the starting point, not the fallback.
AI education and internal capacity building
The FIA didn’t just invest in AI tools – they invested in their people. The tax authority ran workshops and training sessions not just for developers or analysts, but for senior officials as well. The goal wasn’t to turn every employee into a data scientist, but to build a widespread understanding of AI’s capabilities and limitations. By educating senior leadership, the agency ensured that AI initiatives were understood, supported, and championed at the highest levels.
This top-down and bottom-up approach meant that even those who don’t directly use AI day-to-day could still speak its language, advocate for its value, and assess its outcomes critically. It also helped dispel common myths about AI, fostering realistic expectations and encouraging thoughtful implementation across departments.
In conclusion…
By prioritising clean, structured data and fostering a culture of collaboration over punishment, Taiwan has created a robust foundation for leveraging artificial intelligence to enhance compliance and combat fraud.
Its innovative use of tools like graph theory, unified invoicing, and receipt lotteries demonstrates how technology and behavioural incentives can work hand in hand to drive efficiency and trust.
For tax authorities worldwide, Taiwan’s journey offers a compelling blueprint: invest in data quality, empower your people, and embrace digital transformation with a focus on support rather than suspicion.
The future of tax compliance is here, and it’s being shaped by those who dare to rethink the possibilities. Invest in data quality, empower your tax team, and embrace digital transformation is also the approach any company should also take if they want to have a successful tax process.
